A Review of Intelligent Trajectory Planning and Optimization for Aerospace Vehicles

Abstract

Aerospace vehicles operate across a wide flight envelope, traversing dense atmospheric layers from near-space to low Earth orbit. Trajectory planning and optimization in a large spatial domain and wide speed range present severe challenges to traditional methods, including computational efficiency, model accuracy, and constraint adaptability. Artificial intelligence provides an effective pathway to overcome these limitations and has become a key driver for advancing trajectory planning and optimization of aerospace vehicles. This paper presents a systematic review of the core characteristics of aerospace trajectory planning, including environment coupling, multi-constraint compliance, propulsion integration, and aerodynamic nonlinearity, as well as the limitations of traditional methods. The study focuses on the application of intelligent algorithms in both the ascent and reentry phases. For the ascent phase, three key issues are addressed: multistage hybrid optimization with continuous and discrete variables, propulsion multimodal–trajectory coupling, and trajectory reconfiguration under engine failure. For the reentry phase, discussions are focused on such technical difficulties as multi-constraint trajectory generation, no-fly zone avoidance, and multi-mission requirement optimization. Finally, future research directions in intelligent trajectory planning and optimization are discussed, providing theoretical support and methodological guidance for the autonomous and intelligent development of aerospace vehicle trajectory planning.

1. Introduction

1.1. Research Background and Significance

The global aerospace industry has entered a new phase of development, characterized by the large-scale deployment of satellite constellations, crewed lunar exploration, and other landmark programs. As the cornerstone of all space activities, space transportation systems are facing a growing demand for reusability, while addressing essential issues in achieving low cost, high launch frequency, and high reliability [1]. Aerospace vehicles integrate advanced aviation and space technologies. These vehicles hold strategic importance in military reconnaissance, rapid global strike capabilities, and space exploration. Trajectory planning and optimization are critical technologies for aerospace vehicles to complete their missions, directly influencing the flexibility, robustness, and optimality of guidance systems. At present, increasing mission complexity and dynamically evolving cross-domain environments demand urgent breakthroughs in traditional aerospace vehicle trajectory planning to address major challenges. These challenges include: cross-domain dynamic coupling (the interactions between aerodynamic effects, propulsion characteristics, and trajectory parameters), multi-objective conflicts (such as fuel consumption, corridor constraints, and terminal accuracy), and significant uncertainties (such as aerodynamic parameter disturbances, thrust deviations, and sudden failures) [2].

The advancement of Artificial Intelligence (AI) technology has introduced a range of innovative solutions to overcome the limitations of conventional methods. AI facilitates the development of an integrated framework for trajectory modeling, optimization, and control. By seamlessly combining data-driven approaches with mechanistic models, it enables end-to-end planning and auxiliary optimization, significantly enhancing the autonomous decision-making capabilities in trajectory generation. Currently, intelligent algorithms such as Reinforcement Learning (RL) and Deep Neural Networks (DNNs) have demonstrated distinct advantages in managing complex constraints and high-dimensional nonlinear optimization [3]. These technologies provide feasible solutions for flight scenarios, including autonomous trajectory planning for aerospace vehicles in cross-domain, multi-constraint environments, and trajectory reconstruction in the event of sudden propulsion failures [4]. This review systematically explores the technological advancements and innovative applications of intelligent planning and optimization methods, establishing a theoretical foundation and providing methodological references for the intelligent autonomy of aerospace vehicle trajectory planning.

1.2. Flight Trajectory Characteristics of Aerospace Vehicles

The trajectory characteristics of aerospace vehicles can be described from four key perspectives: environmental coupling, multi-constraint compliance, propulsion integration, and aerodynamic nonlinearity. Examples of such cross-domain aerospace vehicles include launch vehicles [5], hypersonic vehicles [6], spaceplanes [7], reentry vehicles [8], marphing aircraft [9], and reusable launch vehicles [10]. These vehicles are designed to operate across a range of environments, from dense atmospheres to near-space and low-Earth orbit [11]. Their flight characteristics are defined by broad speed regimes, extensive airspace coverage, and long-range capabilities, which result in trajectory properties that distinguish them from conventional aircraft [6,12].

First and foremost, trajectory modeling is significantly influenced by multi-physics effects in complex, cross-domain environments. Trajectory planning demands a seamless transition from dense atmospheric conditions to near-space. This transition causes abrupt, orders-of-magnitude changes in environmental loads, such as aerodynamic forces and thermal effects on the vehicle. Due to their typically slender fuselages, aerospace vehicles are subjected to substantial normal aerodynamic loads under wind fields, which may lead to structural failure [10]. In environments involving rarefied gas effects, high-temperature gas effects, and flow transition phenomena, aerodynamic forces and moments across multiple channels exhibit strong coupling, and high nonlinearity [13]. These challenging characteristics demand that aerodynamic modeling be highly adaptable to extreme environmental gradients.

Secondly, trajectory planning exhibits inherent coordination characteristics with multi-modal propulsion systems [14,15]. This provides strong acceleration capability, wide-range operational capability, and excellent economic efficiency for aerospace flight. Its mainstream variants include the Turbine-Based Combined Cycle (TBCC) [16] and Rocket-Based Combined Cycle (RBCC) [17]. The operation and transition between multiple propulsion modes (turbine, ramjet, rocket) in combined-propulsion aerospace vehicles exhibit strong correlation with flight trajectories. Trajectory planning must be deeply integrated with the operational boundaries and transition logic of different propulsion modes [7]. The trajectory parameters directly influence the available operational mode of the power system, and the switching of propulsion modes involves a gradual transition from the pre-state to the post-state. During this process, the engine inevitably experiences a significant thrust reduction, which can lead to aircraft stall and prevent successful initiation of the post-state mode. Furthermore, engine parameters are highly sensitive to changes in aircraft attitude during the mode transition, significantly enhancing the coupling between the aerodynamic and propulsion systems. As a result, these systems form a closely interdependent, bidirectional cooperative relationship. Therefore, trajectory design must precisely coordinate the dynamic adjustments of the aircraft’s intake and exhaust systems, internal flow channels, combustion control, and attitude control systems, optimizing the parameter evolution during the mode transition phase.

Thirdly, trajectory planning and optimization design are subject to stringent multi-dimensional constraints. Strict adherence to path constraints including dynamic pressure, heat flux, and overload is essential during cross-domain flight and air-breathing engine operation. These constraints ensure efficient intake and compliance with several critical factors: angle of attack limits derived from engine total pressure recovery, thrust limitations due to flameout and surge margins, mode transition junction constraints, and orbit injection point restrictions. Whether in ascent to orbit or during reentry, multi-dimensional terminal state constraints must be considered. Additionally, mission-specific environments impose varying constraints on flight parameters like angle of attack and sideslip angle, which may cause these parameters to exceed their prescribed limits, thereby degrading engine performance. Moreover, fluctuations in the internal flow field introduce stricter requirements on the dynamic variations of the angle of attack and sideslip angle. The constraints across different mission scenarios collectively define a narrow flight corridor, placing exceptionally high demands on the accuracy and robustness of trajectory planning.

Lastly, not only the aerodynamic coefficients are highly nonlinear, but the thrust coefficient expressions of the multi-mode engine also exhibit significant nonlinearity and strong coupling. Aerospace vehicle dynamic models are inherently complex, multidimensional, and coupled nonlinear systems. Under the integrated airframe/propulsion characteristics [18], there exists a non-negligible mutual coupling between attitude motion and trajectory variables, as well as between aerodynamic and propulsive effects [19]. The control variables include the angle of attack, sideslip angle, throttle setting, and fuel consumption rate. These design variables interact to affect both dynamic characteristics and engine performance. Additionally, the system is influenced by disturbances in aerodynamic and attitude parameters, further exacerbating the nonlinearity of the flight dynamics model. In such conditions, both the translational motion of the center of mass and the attitude motion around it govern the controllability of trajectory planning. This necessitates solving trajectory generation and flight control as a coupled, integrated problem.

In summary, the trajectory planning and optimization of aerospace vehicles are shaped by the combined effects of cross-domain flight environments, multidimensional complex constraints, multimodal combined propulsion systems, and their inherent nonlinear characteristics. These fundamental factors collectively define the unique complexity and challenges associated with their trajectory planning and optimization.

2. Limitations of Traditional Trajectory Planning and Optimization Methods

The process of solving aircraft trajectory planning and optimization problems can generally be divided into four steps: mathematical modeling of system dynamics, definition of performance metrics, development of solution methodologies, and implementation of corresponding strategies [20]. However, for aerospace vehicles, multiple factors influence their performance during cross-domain flight missions. These include mission requirements, target accuracy, flight environment, path constraints, and other variables, each exerting varying influences on the aforementioned steps. As a result, the inherent limitations of traditional trajectory planning methods have become increasingly evident when addressing the multi-domain, highly dynamic nature of aerospace vehicle flight missions [21].

2.1. Real-Time Trajectory Planning Computational Efficiency Bottlenecks

Existing trajectory planning and optimization methods often adopt indirect or direct approaches. While these methods can achieve nonlinear optimal solutions, real-time computational efficiency remains a significant bottleneck [22]. Major concerns include sensitivity to initial guesses, which can easily lead to convergence difficulties, as well as excessively long computational times for online optimization. Moreover, solution stability across sequential iterations is often insufficient [23]. Aerospace vehicle trajectory optimization is inherently a high-dimensional, strongly nonconvex optimization problem [24]. Traditional numerical iterative or convex optimization-based methods reformulate the original problem equivalently as a convex problem through appropriate variable transformations and precise convex relaxations. However, when initial guesses are poor, methods such as Sequential Convex Programming (SCP) may struggle to converge [25]. This issue referred to as artificial infeasibility, arises from inappropriate convexification. It oversimplifies the constraint set, leading to a lack of intersection between the trust region and the convexified constraint set [26]. Path planning algorithms are mainly used to generate a geometric path from an initial point to a target point, passing through predefined via-points [27]. The path can be defined in either the joint space or the operational space. Trajectory planning algorithms, on the other hand, assign time information to a given geometric path, thereby forming a complete motion description. However, due to their limited adaptability, their performance degrades significantly in dynamic and uncertain environments with complex constraints. Additionally, artificial infeasibility can cause oscillatory behavior that hinders convergence within the trust region, resulting in state fluctuations in the objective function [28]. These challenges not only make it difficult to achieve rapid responses with limited computational resources, but also increase sensitivity to initial guesses. Consequently, they fundamentally constrain the ability to perform real-time online trajectory planning and optimization [29].

2.2. Limitations in the Fidelity of Kinetic Mechanism Models

Model fidelity represents the degree to which a mathematical model can accurately replicate the dynamic behavior of an actual physical system [30]. Existing trajectory planning and optimization methods often employ low-precision dynamic modeling approaches, treating the aircraft as a point mass to describe its flight path. The aerodynamic data used for flight trajectory calculations are based on empirical engineering values, which remain fixed during the flight. This neglects the interaction between aerodynamics and trajectory, which can result in deviations, even if the simulated trajectory closely matches the actual flight path [31]. Additionally, the three-degree-of-freedom center-of-mass dynamics equations can describe the flight position and velocity, but the generated trajectory is difficult to achieve in actual physical systems [32]. In practice, aerodynamic data undergoes real-time variation with changes in altitude and speed, resulting in the consequent alteration of trajectory calculations. The root cause is that physical mechanism models cannot accurately capture the complex dynamics of real systems, leading to inconsistencies between planning results and actual flight environments.

Traditional methods heavily rely on simplified aircraft dynamics models derived from physical equations. However, when addressing strong nonlinear aerodynamic effects during cross-domain flight, switching between multi-modal propulsion systems, and dealing with uncertain atmospheric conditions, the fidelity of such mechanism-based models significantly deteriorates [33]. The fixed-parameter nature of these models prevents dynamic adjustment of parameters, particularly in the event of sudden propulsion malfunctions, actuator failures, or environmental disturbances. In convex optimization approaches, the solution process involves iteratively using a trust region constraint around the reference point to obtain local solutions, gradually approaching the optimal solution. Nonetheless, this approach becomes problematic when the aircraft state deviates from the reference point, resulting in overly conservative or even infeasible planning outcomes. The mismatch between the physical mechanism model and actual dynamics leads directly to inaccurate trajectory planning, poor optimization, and distorted constraint satisfaction, thereby jeopardizing mission safety and reliability.

2.3. Complex Constraints and Multi-Objective Optimization Challenges

Existing methods for trajectory planning of aerospace vehicles exhibit limitations in handling complex, multidimensional constraints and lack adaptive coordination capabilities in dynamic environments. During aerospace flight operations, it is essential to satisfy simultaneously stringent path constraints, terminal constraints, and potentially sudden state constraints (e.g., fairing jettison triggered by thermal flow limitations). These constraints are highly complex and challenging to convexify [34].

Traditional methods can expand the feasible region using slack variable techniques. But they struggle to dynamically balance the priorities and interdependencies between constraints [35]. In trajectory planning, segmenting constraints based on velocity poses challenges, as it complicates the direct handling of relevant constraints. The issue arises because the constraints are defined in terms of velocity-based segments, yet velocity itself is an unknown state variable. Traditional approaches introduce segmented time nodes to optimize variables and velocity constraints, but transforming the original single-stage fixed-end-time problem and increasing its inherent complexity [36].

As mission complexity increases, existing deterministic optimization methods for single objectives are no longer adequate to meet the engineering design requirements associated with complex system characteristics and multi-task, multi-objective collaborative optimization [37,38]. Multi-objective optimization involves balancing several metrics, such as fuel consumption, time efficiency, and thermal load, while also addressing the challenge of multiple local optima arising from the interaction of various tasks [39]. In this review, the concept of a global optimum is replaced by the Pareto-optimal front, which represents the trade-offs between competing objectives through a set of non-dominated solutions [40]. The multi-objective complexity of aerospace vehicles requires optimizing one objective through interactive decision-making without compromising others. However, a notable gap exists in the availability of data-driven “Pareto optimal frontier” solution sets for trajectory selection.

2.4. Comparative Analysis of Traditional and Intelligent Trajectory Planning Methods

It can be observed that traditional methods face significant limitations in many contemporary application scenarios.

Table 1. Comparison Between Traditional and Intelligent Methods in Different Scenarios.

Scenario	Traditional & Formula	Intelligent & Formula
Autonomous Entry Trajectory Planning	SOCP [41] $x_{N+1}=x_0+{\displaystyle \frac{{\overline{e}}_f-{\overline{e}}_0}{2}}{\sum }_{m=1}^N{w}_m\left[A{x}_m+f_3{u}_m+b\right]$ (1)	RL [41] $y=r+\gamma {\min }_{i=1,2}{Q}_i^{\prime }\left(s^{\prime },k^{\prime },{\tilde{x}}_{k^{\prime }}|{\theta }^{Q^{\prime }},{\theta }^{A^{\prime }}\right)$ (2)
Real-time ReentryTrajectory Planningfor HypersonicVehicles	RPM [42] ${\displaystyle \frac{\mathrm{d}x}{\mathrm{d}\tau }}={\sum }_{k=0}^{N_k}x\left({\tau }_k\right){\displaystyle \frac{\mathrm{d}{L}_k}{\mathrm{d}\tau }}={\sum }_{k=0}^{N_k}x\left({\tau }_k\right)D_{jk}\left({\tau }_k\right)$ (3)	DNNs [42] $a^l=\sigma \left(z^l\right)=\sigma \left(W^l{a}^{l-1}+b^l\right)$ (4)
Skip–glide trajectory scenario	Function Approximation [43] $\dot{x}=f(x,u,t)$   (5)	LSTM [43] $\begin{array}{c}[{\delta }_{k,\mathrm{LSTM}},C_{k,\mathrm{LSTM}}]=\\ \mathrm{LSTM}({\delta }_{k-1,\mathrm{LSTM}},C_{k-1,\mathrm{LSTM}},{\mu }_k)\end{array}$ (6)

summarizes the core formulations of both traditional and intelligent approaches across three representative scenarios, along with a comparative analysis based on existing experimental results. In the scenario of autonomous entry trajectory planning, Peng et al. [41] employs RL for trajectory planning and optimization Equation (2), compared with the traditional Second-Order Cone Programming (SOCP) based method Equation (1). The proposed approach achieves a model accuracy of 93.55%. In terms of adaptability, it demonstrates high terminal precision and maneuver efficiency, while robustness is improved by 19.1%. In the scenario of real-time reentry trajectory planning, Peng et al. [42] utilizes the DNNs Equation (4) to generate optimal trajectories from training samples. Compared with the Radau Pseudospectral Method (RPM) Equation (3), the position and velocity errors are reduced by 59% and 86%, respectively. Meanwhile, the computational speed is improved by 67 times, achieving microsecond-level command output. This ensures optimality while satisfying the stringent requirements of hypersonic vehicle reentry, including strong real-time performance, high precision, and robustness. For the skip–glide trajectory scenario, Li et al. [43] demonstrates that the use of Long Short-Term Memory (LSTM) Equation (6) significantly reduces trajectory prediction errors compared to analytical fitting models Equation (5), while requiring less prior knowledge.

Therefore, improving computational efficiency, model accuracy, and constraint adaptability is essential. It helps achieve autonomy, real-time performance, and reliability in aerospace vehicles. This makes the application of intelligent methods to trajectory planning an inevitable trend.

3. Intelligent Trajectory Planning and Optimization Method

Intelligent trajectory planning methods primarily include bionic heuristic algorithms, RL, and others. Bionic heuristic algorithms serve as the foundation for intelligent trajectory planning and optimization. These algorithms have fewer hyperparameters and do not require a training process, enabling them to generate approximately optimal trajectories in a short amount of time. For example, Genetic Algorithm (GA) [44], Ant Colony Optimization (ACO) [45], Whale Optimization Algorithm (WOA) [46] and Particle Swarm Optimization (PSO) [47] have gained significant attention for addressing constrained trajectory planning problems. Heuristic optimization algorithms do not depend on gradient information or prior knowledge of feasible solutions, which often allows them to identify global optimal solutions [48].

Recent studies have increasingly focused on the application of data-driven and machine learning techniques in trajectory planning and optimization [2,49,50,51]. A significant connection exists between machine learning and evolutionary algorithms, particularly in the generation of training data. Once machine learning models are used to generate high-quality initial guesses, the performance of subsequent evolutionary operations can be notably enhanced. Among these methods, DNNs have demonstrated superior performance [52]. Research indicates that, in the context of unmanned aerial vehicles and spacecraft, DNNs architectures can generalize well beyond the boundaries of the training dataset. However, the acquisition of reliable data from space remains one of the most significant challenges in the field of Deep Learning (DL) [53].

As a branch of machine learning, RL aims to train agents to learn effective trajectory planning strategies in dynamically changing environments. The system obtains feedback through the reward function, which can evaluate the current performance of the agent in real time. The advantage of RL is that it can adapt to unpredictable unknown environments. Its objective is to maximize the cumulative reward by selecting an optimal sequence of actions based on observations of the environment. Hua et al. [54] proposed a learning-based framework for aircraft trajectory generation. It integrates RL and imitation learning to facilitate intelligent decision-making in real-time, ensuring the practicality and timeliness of the solution.

Trajectory planning and optimization is essentially an optimal control problem. A prominent intersection between AI and optimal control is the application of AI techniques to the optimal trajectory design of aerospace vehicles [37]. Unlike traditional trajectory optimization methods [55,56,57,58], the application of AI based approaches has significantly shortened the design cycle and reduced reliance on specialized knowledge. In turn, this has increased the likelihood of identifying the global optimal solution for the task. Li et al. [59] employed a trained AI model to approximate the optimal multi-impulse transfer trajectory as well as the corresponding cost for multi-objective problems. This approach is particularly valuable for tasks that require rapid estimation of optimal trajectories or calculation of performance indices.

In many aerospace missions, atmospheric entry or exit is a crucial component. These missions span a wide range of aerospace vehicles, including space shuttles, launch vehicles, ballistic missiles, reusable launch vehicles, and hypersonic vehicles [60]. During ascent and reentry phases, flight trajectories typically span large spatial domains and extended flight envelopes over prolonged durations, making trajectory planning and optimization essential to mission success.

3.1. Intelligent Planning and Optimization of Ascent Trajectory

The ascent phase represents the initial stage of space mission execution. During flight profile design, it is essential to comprehensively account for the vehicle model and various constraints in accordance with mission requirements, so as to generate optimal trajectories tailored to different mission objectives [61,62,63]. In addition to optimizing trajectory variables, thrust vectoring can also be incorporated as a decision variable. This can further enhance ascent performance while satisfying fuel efficiency requirements and multiple operational constraints [64]. These constraints, including initial conditions, terminal conditions, and path constraints, are formulated within the framework of an optimal control problem [65]. As shown in Figure 1, research on intelligent ascent trajectory planning and optimization exhibits both typical solution paradigms and key technical challenges. The figure presents two representative launch modes: vertical takeoff and horizontal takeoff. For vertical takeoff vehicles, trajectory planning involves several critical flight events. These include stage separation, fairing jettison, and payload separation. For horizontal takeoff vehicles, the propulsion system operates in multiple modes. Smooth and reliable mode transitions are required. To address these challenges, three aspects are considered. The first is multistage hybrid optimization with continuous and discrete variables. The second is propulsion multimodal–trajectory coupling. The third is reconfiguration planning under engine power failure.

• Multistage Continuous and Discrete Variable Hybrid

The hybridization of continuous and discrete variables is a significant feature of ascent flight [66]. The ascent phase of aerospace vehicles typically consists of several phases, including the boost phase, climb/cruise phase, and load separation phase. During these phases, the vehicle faces continuous mass changes due to fuel consumption, as well as discrete mass changes caused by stage separations and fairing ejection. Additionally, the discrete state variables resulting from propulsion mode transitions, particularly in vehicles with combined propulsion systems, further complicate the dynamics. This introduces hybrid continuous/discrete characteristics into the aircraft dynamics, which increases the complexity of trajectory planning and optimization by incorporating various data types and model scales.

With the successful application of DL in control optimization, researchers have leveraged the strong fitting capability of DNNs to model nonlinear systems. These models are further applied to predict and approximate hybrid trajectories of aerospace vehicles that involve both continuous and discrete variables [67]. Yang et al. [68] noted that a single DNN model may struggle to adapt to the multi-stage dynamic variations of rockets. To address this, they significantly improved the model’s adaptability to complex dynamics through hierarchical, segmented training and multi-model fusion frameworks tailored to different data types. They also introduced a knowledge-driven strategy, which incorporates physical models or expert knowledge to guide network training. To reduce training costs and shorten learning times, the design of an intensive reward mechanism was implemented, substantially enhancing training efficiency [69]. To address the issue of continuous switching between booster rocket flight modes and the associated complex dynamics, Zang et al. [70] developed a model set combining the gravity steering model, three-dimensional steering model, and a precise dynamic model. Convolutional Neural Networks (CNNs) were employed to extract feature information from the target state, while a compressed excitation network was introduced to accurately distribute the importance of CNN feature channels. Additionally, LSTM networks were utilized to enable real-time updates of model probabilities, improving the model switching speed. Although these methods successfully characterize the continuous/discrete hybrid characteristics in aircraft trajectory planning, they do not fully address trajectory generation across multiple data types and model scales.

• Propulsion Multimodal and Trajectory Coupling

Ascent trajectory planning faces the central challenge of integrating multi-mode propulsion systems with the coupling between aerodynamic and propulsion. Traditional numerical optimization methods typically treat trajectory planning and propulsion optimization as separate tasks. This separation leads to insufficient coordination between the two, significantly limiting the overall performance of the aerospace vehicle [71]. Therefore, it is crucial for ascent trajectory planning to be closely integrated with engine characteristics [72].

During the ascent phase of aerospace vehicles, the vehicle typically climbs with constant dynamic pressure to ensure optimal or stable operating conditions for the propulsion system. Due to the strong interdependence between engine performance and flight trajectory, a coupling exists between trajectory design and engine performance optimization. Zhang et al. [73] proposed an online trajectory (velocity–dynamic pressure rising trajectory) based on the operational boundaries of the engine under stringent constraints, utilizing a preset altitude tracking method to achieve the ascent profile. Building on the development of an engine component-level model, Zhang et al. [74] introduced a combined engine mode conversion control scheme after determining the mode conversion working point. They designed a multi-variable controller based on neural network estimation and inverse control, enabling simultaneous trajectory planning and engine control. However, the lack of dynamic adaptability in the mode conversion working point limits the robustness of trajectory planning and optimization, particularly under uncertain disturbances and extreme conditions. To address these uncertainties, such as wind field fluctuations, thrust deviations, and other disturbances, Zhou et al. [75] constructed a multi-scenario disturbance database and employed neural network algorithms to reconstruct the flight profile in real time, overcoming the reliance on traditional guidance systems that depend on precise disturbance models. By integrating a hybrid algorithm strategy for global search and local optimization, they significantly improved the efficiency of solving complex constrained problems. However, the strong coupling between airframe and propulsion systems, inherent in the airframe/propulsion integration, results in extremely high modeling complexity, low solution efficiency, and a heavy reliance on data. Currently, trajectory planning and optimization for aerospace vehicles under multi-modal propulsion characteristics are typically based on segmented designs, which struggle to account for dynamic uncertainties, mode transitions, and the integration of airframe/propulsion systems in a unified optimization framework.

• Reconstruction Planning under Engine Power Failure

In the history of aerospace, failures during vehicle launches caused by thrust drop are not uncommon [76]. The robustness and re-planning capabilities of ascent trajectory optimization are critical for meeting mission requirements, particularly for injecting the payload into the target trajectory. Ascent trajectory re-planning after failure is typically performed in two steps: the first step involves determining the reachable and safe target trajectory based on the residual capacity of the launch vehicle. The second step is to generate the ascent trajectory in real time [77]. The uncertainty surrounding the timing and extent of the failure significantly increases the complexity of real-time trajectory reconstruction. Therefore, Mission Reconstruction (MRC) planning during the ascent phase is crucial to minimizing the loss caused by power failure.

Unlike traditional ascent trajectory optimization, which uses a specified target trajectory, MRC requires real-time optimization of the target parameters [78]. In the event of failure, when the target needs to be altered and the mission cannot be interrupted in flight, autonomous MRC must be urgently implemented to salvage the mission. This introduces significant technical challenges for online optimization [79]. To address this, a large number of degraded trajectories and trajectory samples were generated using Nonlinear Programming (NLP) solvers to train the DNNs offline [80]. The trained network provides accurate target trajectory and initial reference trajectory information, thereby improving the efficiency of online trajectory optimization algorithms.To solve the problem of trajectory re-planning after failure, a suboptimal joint trajectory re-planning method was proposed in the literature [81]. By employing a learning-based warm-up startup scheme, the DNNs are trained offline to provide a reasonable initial guess, preventing infeasible solutions and improving solution efficiency. When both rescue trajectory and flight trajectory optimization problems arise simultaneously, traditional optimization methods are time-consuming. To address this, He et al. [82] introduced a fault rescue knowledge integration model based on a Probabilistic Neural Network (PNN) and Radial Basis Function Neural Network (RBFNN) to determine the optimal rescue trajectory. This approach provides terminal constraints for the trajectory optimization problem through training, effectively reducing the search space for the optimal solution.

3.2. Intelligent Planning and Optimization of Reentry Trajectory

The trajectory planning and optimization of the reentry phase of aerospace vehicles present significant challenges due to highly nonlinear dynamics, multiple constraints (such as terminal state, no-fly zones, and heat flux limits), and the need for real-time computation [83,84]. Reentry vehicles are particularly influenced by state uncertainties arising from both internal and external sources. Internal uncertainties include measurement errors and sensor noise, which are introduced into the system, while external uncertainties are generated outside the vehicle and directly affect its flight state [85]. Traditional numerical optimization methods often struggle to meet the terminal accuracy requirements when dealing with such complex constraints. This highlights the urgent need for intelligent algorithms that balance both computational efficiency and optimization accuracy [86].

Figure 2 illustrates the reentry flight scenario of aerospace vehicles from near-space to the ground. Three representative tasks are summarized. The first is vertical recovery. It involves trajectory planning from the reentry interface to attitude adjustment, and then to horizontal recovery. The second is horizontal landing. The vehicle passes through the reentry corridor, avoids no-fly zones, and completes trajectory planning for horizontal landing. The third is atmospheric entry–exit trajectory planning. The vehicle performs brief entry and exit maneuvers based on on-orbit information acquisition. The following sections will discuss these aspects in detail.

• Trajectory Generation under Multiple Reentry Constraints

The reentry flight trajectory is a time series with nonlinear characteristics, which can be effectively approximated by neural networks with strong nonlinear approximation capabilities. The remaining flight trajectory can be predicted using DNNs [87]. Leveraging advancements in AI algorithms, some studies have proposed combining computational intelligence with control theory to enable fault-tolerant or self-healing control systems, thereby alleviating the design burden [88]. Given the time-dependent nature of trajectory data, Recurrent Neural Networks (RNNs), particularly LSTM networks, have been widely used for trajectory optimization. In [43], the authors demonstrate that the height, velocity, and velocity direction of the reentry terminal can be expressed as analytical functions. By analyzing the dynamic characteristics of model parameters, they propose a target trajectory prediction channel that incorporates prediction errors. Additionally, an improved LSTM network is designed to dynamically update and approximate trajectory parameters. In [89], the Extreme Learning Machine (ELM) is employed to predict the reentry trajectory, meeting the stringent real-time requirements of reentry trajectory optimization. To enhance the generalization performance of the ELM, a regularization term is introduced, and the Marine Optimization Algorithm (MOA) is utilized to optimize network parameters and accelerate convergence.

To meet the complex multi-constraint conditions (dynamic pressure, heat flow, overload constraints) of the reentry corridor and ensure the convergence of the optimization problem under these constraints, Li et al. [90] employed a double-delay deep deterministic strategy gradient combined with a continuous convex optimization method. This approach allows for online adjustment of the performance index function, optimizing the trajectory configuration. To address the sparse reward problem caused by multiple constraints, Jiang et al. [91] proposed an improved ex post experience replay method, which is adaptively integrated with the Deep Deterministic Policy Gradient (DDPG) algorithm by transforming multiple constraints into multiple objectives. Furthermore, the combination of feedforward neural networks and the pseudospectral discretization method within convex programming significantly reduces the solution time for multi-constrained glide trajectory optimization [92]. In addition, for reentry terminal angle and trajectory constraints, Tong et al. [93] introduced a roll angle predictor based on a Transformer network during the glide phase, enabling precise fine-tuning of the heading angle under the soft constraints of reentry velocity. In the final reentry phase, an improved Beluga Whale Optimization Algorithm is proposed to correct errors through feedback for the nonlinear, fast optimization problem under complex multi-constraints.

• Trajectory Planning for Reentry No-Fly Zone Avoidance

A no-fly zone is a geographical area that must be avoided due to air defense threats and political restrictions. The reentry trajectory, which intersects such no-fly zones, must account for complex spatial constraints. Common methods for no-fly zone avoidance include energy-optimal waypoint planning [94], rollover logic avoidance [95], and the artificial potential field method [96], among others. However, these methods often suffer from low computational efficiency. Additionally, when the no-fly zone changes, accurately re-planning the trajectory becomes a challenging task [97]. Search-based solutions, a cornerstone of early AI algorithms [98], offer an alternative approach. The A-Star search algorithm, widely used in graph search, demonstrates efficient search performance by exploring feasible paths through an evaluation function. As one of the key algorithms for solving the shortest path search problem, the A-Star search algorithm can perform in-depth exploration based on spatial information [99].

RL does not rely on prior models. When the environment is unknown and dynamic, RL demonstrates better robustness to uncertainty compared to traditional algorithms [100]. Zhou et al. [101,102] developed a hybrid optimization framework combining PSO and Deep Reinforcement Learning (DRL) for Reusable Launch Vehicles. In this framework, DRL dynamically adjusts hyperparameters based on particle swarm behavior and adaptively controls particle motion and evolution. The reward function is redefined by incorporating a maximum entropy term and adaptive noise, which further enhances the exploration of the global optimal solution [103]. However, model-free RL often requires frequent interactions, leading to high computational costs, especially when working with accurate models. To address this, Li et al. [104] adopted a prediction-correction framework trained via Model-Based Policy Optimization (MBPO) and tackled the multi-local optimal solution problem in trajectory optimization through branch-rolling. This approach significantly improves the training efficiency of trajectory planning strategies compared to Proximal Policy Optimization (PPO). Wu et al. [105] proposed a method based on the naturally inspired Interference Fluid Dynamics System (IFDS), which enables comprehensive avoidance of multiple no-fly zones and unified judgment of avoidance rules. To improve adaptability and robustness, a DRL optimization method based on the Double-delay Deep Deterministic Policy Gradient (TD3) algorithm is used to enable online adaptive adjustment.

• Trajectory optimization for Reentry Multi-Task Requirements

Trajectory planning during the reentry phase often involves multi-tasking requirements. After the reentry vehicle enters the atmosphere to perform observation or information collection tasks, it may return to near-space to continue its mission. In this phase, the coexistence of multi-tasking requirements, such as re-entering the atmosphere and returning to space, causes conflicts among objective functions in trajectory planning. For example, trajectory designs aimed at reducing fuel consumption may inadvertently increase kinetic energy or total heat flux. To address multi-objective and multi-stage optimization problems, Bae et al. [106] introduced Fuzzy Satisfactory Goal Programming (FSGP) into the multiphase hp-adaptive pseudospectral method. This approach enables a balance between optimality and priority by prioritizing targets in advance and integrating conflicting objectives into a unified objective function. Compared to methods that train optimal strategies through environmental interaction, fuzzy logic can effectively handle complex, uncertain systems that are difficult to model mathematically, especially when data is limited. This decision-making approach, based on fuzzy rules and reasoning mechanisms, also makes the decision process interpretable. In the context of uncertain constraints and dynamics, Chai et al. [107] proposed an interactive fuzzy physical programming method to address reentry trajectory oscillations. However, fuzzy decision-making typically relies on a comprehensive set of expert-defined rules, which can be limiting. To overcome this, Chai et al. [108] designed an integrated framework that combines multi-objective trajectory optimization with DNNs. In this framework, the fuzzy multi-objective transcription method is first used to generate optimal trajectories, and these trajectories are then used to train the DNNs, allowing it to generate real-time optimal guidance.

The trajectory planning and optimization of aerospace vehicles often involve urgent tasks, requiring decisions on multiple objectives to select the most appropriate one. In addition to the need for rapid execution, reentry trajectory generation algorithms must also cover as many task areas as possible. Furthermore, when constructing a RL environment, discretization is necessary to convert a continuous problem into a sequential decision-making problem with finite dimensions. However, the choice of discrete step size can lead to inconsistent states and a trade-off between the accuracy and complexity of network learning. To address these challenges, Peng et al. [42] proposed an adaptive, data-driven reentry trajectory generation algorithm based on the DDPG algorithm, which can adapt to a wide range of complex tasks.

Remark 1.

Since various intelligent methods have been applied to trajectory planning and optimization, it is necessary to provide a clear and systematic comparison. This helps readers select appropriate methods for their specific applications.

Table 2. Advantages and Disadvantages of Main Intelligent Algorithms for Aerospace Vehicle Trajectory Planning and Optimization.

Method	Advantage	Disadvantage
DNNs [67]	Enhances solution efficiency under complex constraints and enables real-time trajectory optimization.	Requires large-scale disturbance datasets with high data acquisition cost.
CNNs + LSTM [70]	Improves model switching speed via feature extraction and real-time probability updating.	Complex architecture with high computational cost.
PNN + RBFNN [82]	Enforces terminal constraints and reduces the optimal search space.	Performance depends on training data coverage; limited generalization with insufficient samples.
TD3 [90]	Ensures convergence under complex constraints via online objective adjustment.	High structural complexity and computational burden.
DDPG [91]	Converts constraints into multi-objective optimization to address sparse rewards.	Trade-off between learning accuracy and model complexity; potential state inconsistency.
Transformer [93]	Improves terminal accuracy via error feedback correction.	Complex structure may reduce computational efficiency.
DRL [101]	Enhances optimization by adaptive hyperparameter tuning of swarm-based methods.	High computational cost; coordination between modules remains challenging.
MBPO [104]	Handles multi-local optima via branched rollouts.	Model bias propagates to policy learning and degrades performance.
FSGP [106]	Supports multi-objective prioritization and conflict balancing.	Relies on expert-defined rules with subjective priority settings.

summarizes the advantages and disadvantages of nine mainstream intelligent algorithms used in aerospace trajectory planning and optimization. These include representative paradigms such as deep learning, reinforcement learning, and hybrid neural networks.

3.3. Hardware Computational Bottlenecks and Engineering Challenges

Although intelligent methods have demonstrated remarkable performance gains in theoretical simulations for trajectory planning and optimization, translating these methods from conceptual validation to practical aerospace deployment encounters formidable hardware computational bottlenecks and engineering constraints. According to the synthesis in Table 2, the practical deployment and engineering implementation of intelligent algorithms, such as DNNs [67] and Transformers, remains constrained by the following critical factors.

Firstly, a primary hurdle lies in the contradiction between the structural complexity of high-performance algorithms and the limited onboard computational resources. Advanced models, such as CNNs + LSTM and Transformers, possess intricate architectures that entail significant computational overhead [70]. Unlike terrestrial high-performance computing clusters, aerospace-grade computing platforms prioritize mission reliability in extreme environments, resulting in relatively constrained clock frequencies and memory bandwidths. The “high structural complexity” associated with certain algorithms in Table 2 often leads to computational latencies that fail to meet the microsecond-level response requirements of real-time flight control. Consequently, achieving an optimal balance between precision and real-time performance through model compression and lightweight design remains an urgent bottleneck in resource-constrained embedded environments.

Secondly, there is an acute dependency on large-scale, high-quality datasets, which exacerbates the “Sim-to-Real” gap. DL and RL methods typically rely on extensive data for effective training. Given the prohibitive costs and scarcity of authentic flight data, existing models are predominantly trained within deterministic simulation environments. However, as discussed in the contexts of ascent and reentry phases, unmodeled dynamics in actual flight including complex aero-thermo-elastic coupling and stochastic environmental perturbations induce a pronounced discrepancy between simulation and reality. This environmental inconsistency may lead to performance degradation or insufficient generalization in models like PNN or RBFNN when encountering off-nominal conditions not represented in the training samples [82].

Finally, the inherent opacity of intelligent decision-making mechanisms poses a major obstacle to rigorous engineering Verification and Validation. In contrast to traditional optimal control methods that provide rigorous stability proofs, many intelligent algorithms struggle to offer deterministic safety boundary guarantees, which fundamentally conflicts with the stringent reliability requirements of aerospace missions. Even hybrid strategies, such as DRL or MBPO, face challenges involving model bias propagation or system coordination [101,104], making it difficult to establish provable safety criteria during critical mission phases such as engine failure recovery or no-fly zone avoidance.

The aforementioned bottlenecks, characterized by the multi-dimensional trade-offs among learning accuracy, model complexity, and real-time performance, underscore the necessity for a paradigm shift. These challenges compel researchers to explore novel developmental paths, including mechanism–data fusion paradigms, online adaptive technologies, and lightweight interpretable models, which will be discussed in detail in the following sections.

4. The Future Development Trend of Intelligent Trajectory Planning and Optimization Methods

With the advancement of AI and big data technologies, the scope of trajectory planning and optimization for aerospace vehicles has expanded, driving the continuous evolution of science and technology [109]. Progress in aircraft technology has led to an increasingly diverse range of application scenarios. Different aircraft control systems make varying decisions in the same environment, resulting in distinct trajectories. In trajectory planning, the errors arising from these differing choices cannot be overlooked. In practical systems, due to challenges such as the high complexity of numerical processing, strong nonlinearity, and significant coupling, it is often difficult to simultaneously satisfy all trajectory planning constraints in real-world scenarios. Intelligent trajectory planning and optimization techniques can enhance trajectory accuracy, making it essential to continually incorporate intelligent methods into trajectory planning and optimization processes.

4.1. Mechanism–Data Fusion and Cross-Domain Migration New Paradigm

The core challenge of data-driven trajectory planning arises from the interplay between parameter-intensive complex networks and the nonlinear transformations inherent in neural networks. Unlike classical mechanism-based models, data-driven models are essentially black-box systems. This lack of transparency not only complicates the verification process required for specific tasks but also makes interpreting the underlying principles of network outputs a challenging and necessary endeavor. Therefore, it is crucial to strengthen the fusion of multiple algorithms, combining neural networks with analytical algorithms to enhance both the interpretability and generalization of intelligent systems. Future research will likely focus on developing a hybrid architecture that integrates “mechanism models” with data-driven approaches. A promising solution is to incorporate Physical Information Neural Networks (PINNs) [110] directly into the neural network architecture by embedding differential equations and boundary conditions into the loss function. This approach integrates fundamental physical laws and domain expertise directly into the network, ensuring that the model’s predictions remain data-driven while adhering to established physical principles. By leveraging PINNs, it becomes possible to maintain consistency with aerospace flight dynamics, thereby reducing dependence on large volumes of diverse training data. Building on this, future research could explore the synergistic combination of PINNs with other advanced machine learning technologies to enhance the reliability of data samples.

Cross-domain and cross-modal knowledge transfer will become a key focus of algorithm fusion in the future. The optimal solution can be controlled by learning a small number of “behind-the-scenes parameters” (referred to as solution hyperparameters), rather than directly learning a high-dimensional state-input mapping. More importantly, the relationship between these parameters and the control strategy is more predictable, making it easier to learn with less training data. Most trajectory generation problems can be formulated as variants of optimal control problems. Consequently, solving for hyperparameters typically involves fully defining the maximum principle state variables of the optimal control strategy, or their combined form. The dynamic characteristics of aerospace vehicles vary significantly across different flight stages (such as ascent and reentry trajectories) and propulsion system modes. However, intelligent algorithms can facilitate knowledge sharing through transfer learning, enabling effective adaptation across these diverse conditions.

4.2. On-Line Autonomous Trajectory Planning Adaptive Technology

When an aerospace vehicle faces a task target update or fails to reach a specified position relative to the preset target, it must quickly adapt to the new initial conditions based on existing knowledge. Since trajectory planning requires multiple state variables and dynamic equations, the time and resource consumption associated with the optimal trajectory generalization process are significantly increased. In the future, cross-domain flight will necessitate online autonomous trajectory planning and adaptive technologies to improve aerospace vehicles’ ability to generate high-precision, reliable trajectories in real time when confronted with new scenarios and tasks. To address the challenge of prolonged training times for planning and optimization under new tasks, meta-learning equips agents with the ability to adapt from limited experience [111,112,113,114]. By framing the trajectory planning problem as a few-sample regression task, agents can learn from a small number of optimal trajectories and quickly adapt to new planning conditions, enabling rapid prediction of the optimal trajectory [115]. Furthermore, the training efficiency of the agent can be enhanced through the use of recurrent network structures [116]. RNNs, due to their feedback connections, can store the temporal evolution of processed data in their internal state. This capability allows RNNs to better specialize their network outputs for specific problems and significantly improve performance in non-Markovian, partially observable, or multi-task environments. When attitude dynamics are taken into account, online stochastic planning with large, factorized state and action spaces can be introduced [117]. This approach estimates the quality of actions by aggregating simulations of the states reached by feasible actions. Compared with searching over specific state–action pairs, it can significantly improve computational efficiency. Moreover, when the performance of stochastic policies can reflect state quality, it is sufficient to guide decision-making.

In general, it is challenging to learn a unified action strategy that can adapt to multiple tasks. To improve adaptability and reduce learning difficulty simultaneously, a modular strategy can be introduced to decompose the trajectory design problem into multiple modules. By utilizing a discrete and continuous mixed hybrid action space for autonomous decision-making, this approach effectively avoids the trade-off between learning exploration difficulty and trajectory accuracy [41]. However, autonomous trajectory planning for large-scale tasks still faces significant challenges, and current intelligent learning methods struggle to achieve effective data fusion and strategy representation when dealing with mixed action space problems.

4.3. Intelligent Control of Dynamic Multimodality and Trajectory Planning

With the continuous increase in the requirements of various tasks, aircraft must achieve higher speeds and longer ranges. To meet these mission demands, it is necessary to optimize the modal transformation and trajectory planning of the engine. In complex flight environments characterized by long flight durations, high speeds, and multiple constraints, mixed feature points often arise during the conversion between dynamic modes. The multi-dimensional, discontinuous dynamic equations associated with these transitions can result in excessive computational overhead, preventing the mission requirements from being met. Trajectory planning under dynamic multi-mode conditions can be achieved by incorporating precise engine control into the trajectory planning process. One approach could involve designing a hybrid control structure that switches between an adaptive neural network controller and a robust controller [118]. This structure can be applied in the design of the aircraft’s control mode, ensuring that the closed-loop system converges to a bounded disturbance set and effectively regulates the vehicle’s flight trajectory. In the trajectory planning of an airframe/propulsion integrated model, it is critical to describe the engine’s operating state quickly and accurately [119]. AI can process the relevant data in real-time, enabling the flight state to be controlled and the engine’s working state to be rapidly described. Notably, the optimal control method using DNNs holds significant potential, particularly for learning nonlinear relationships under different modes to solve trajectory optimization problems. The constructed DNN model is optimized through various parameters, including the number of layers, neurons, learning rate, and activation function, in order to solve the multi-modal optimal trajectory problem. This three-dimensional real-time trajectory optimization method based on DNN models is expected to achieve near-optimal control with real-time performance and stable convergence [120]. DNNs have the capability to approximate any nonlinear system [121], continuously and accurately mapping inputs, outputs, and the optimization model in optimal control. However, their dependence on large sample sets means that future work will need to focus on continuously improving optimization methods.

Currently, propulsion multi-modal and intelligent trajectory planning, along with optimization and control, face several challenges, including the accurate modeling of multi-modal nonlinear coupling systems, the domain offset between simulation and real environments, and the difficulties in real-time collaborative optimization and constraint satisfaction. In the future, intelligent control of propulsion multi-modality and trajectory planning can be achieved through approaches such as designing lightweight, interpretable DNN models, constructing a multi-modal collaborative adaptive control framework, and combining multi-objective dynamic decision-making with robust optimization.

4.4. Intelligent Planning of Cluster Collaboration and Distributed Trajectory

In mission completion, each aircraft must not only plan its trajectory independently but also cooperate with other aircraft to avoid resource waste and communication bottlenecks [122,123]. Centralized control can enhance decision-making quality. A class of RL models based on proximal policy optimization, cloud-based control systems, and other AI integrated methods can be designed [124]. A single central controller can plan the distributed flight trajectories for all aircraft. However, as the number of aircraft increases, challenges arise related to communication networks, real-time response capabilities, and computational resources. Therefore, focusing on a centralized cluster aircraft model, which utilizes local information for efficient and low-latency control, becomes a viable approach [125,126]. The IoT framework can provide effective technical support for distributed cluster systems [127]. By incorporating distributed sensing, cluster-based information exchange, and autonomous trajectory planning mechanisms, this architecture has the potential to alleviate communication pressure, improve real-time performance, and facilitate efficient cooperative trajectory planning among multiple aircraft. Despite its potential, the high complexity of the algorithm and the difficulty of coordinated optimization limit the performance of tasks such as trajectory coordination optimization. In this context, the Symbolic Aggregate Approximation (SAX) [128] method can be employed to discretize time series data into symbolic sequences. By reducing the dimensionality of the time series, it can significantly decrease data volume and thereby improve computational efficiency. The no-fly zone is a core hard constraint in trajectory planning. Multi-agent reinforcement learning (MARL) has been widely adopted for aircraft cluster collaboration. In the future, various complex AI control methods based on the MARL framework [129] can be designed to enable aircraft to avoid no-fly zones and achieve distributed trajectory planning through autonomous interaction, trial-and-error learning, and collaborative decision-making. In scenarios with scarce data, high real-time demands, and stringent constraints, imitation learning offers a promising solution. It complements RL and meta-learning, and the hybrid imitation RL framework could further enhance the performance of RL.

The intelligent planning of aerospace vehicle cluster collaboration and distributed trajectory optimization holds significant promise. However, existing AI methods still exhibit notable shortcomings in terms of constraint model accuracy, collaborative planning efficiency, and environmental adaptation. In the future, research should focus on optimizing algorithms, distributed training optimization, and other related areas. One promising approach is the use of multi-agent DRL to explore intelligent cooperative control strategies in nonlinear environments. Furthermore, it is essential to consider additional constraints such as communication delays and collision avoidance. Designing intelligent cooperative control laws that can handle complex scenarios will be key to achieving global optimization in cooperative decision-making.

5. Conclusions

Aerospace vehicles are the core carriers for low-cost, high-frequency, and highly reliable space transportation systems. The performance of trajectory planning and optimization directly determines the flexibility, robustness, and optimality of mission execution. This paper systematically reviews existing research on intelligent trajectory planning and optimization for aerospace vehicles. It clarifies the key challenges in both ascent and reentry phases. It also points out current limitations. Ascent trajectory planning is mainly constrained by three major bottlenecks. These bottlenecks include the hybrid nature of multistage continuous and discrete variables, the strong coupling between multimodal propulsion and trajectory, and trajectory reconfiguration under dynamic fault conditions. In the reentry phase, intelligent methods are required to address multi-constraint adaptation, no-fly zone avoidance, and multi-mission trade-offs. The goal is to improve the accuracy, efficiency, and robustness of trajectory generation. Intelligent algorithms such as deep learning and reinforcement learning have shown significant advantages. They have effectively compensated for the limitations of traditional methods. However, several challenges remain. First, mechanism–data fusion models lack sufficient interpretability and generalization. Second, cross-domain transfer capability is limited. Third, online autonomous adaptive methods have inadequate real-time performance and high precision. Fourth, data dependence constrains the coordination between multimodal propulsion and trajectory. Fifth, cluster cooperation and distributed planning have shortcomings in constraint modeling accuracy and coordination efficiency.

Future research should focus on four core directions. These directions are new paradigms for mechanism–data fusion and cross-domain transfer, online autonomous and adaptive trajectory planning, dynamic multimodal intelligent control and trajectory planning, and cluster-based cooperative and distributed trajectory planning. Existing bottlenecks can be overcome by developing lightweight and interpretable intelligent models. Improving distributed training and cooperative optimization algorithms also helps. In addition, integrating multidisciplinary theories and technologies is essential. This will provide solid theoretical support and technical guidance for the autonomous and intelligent development of aerospace trajectory planning. It will further advance the aerospace industry to a higher level.